1 / 47

CS626: NLP, Speech and the Web

CS626: NLP, Speech and the Web. Pushpak Bhattacharyya CSE Dept., IIT Bombay Lecture 6, 7, 8, 9: Viterbi ; forward and backward; Baum Welch; IL POS tags 6 th , 8 th , 9 th , 13 th August, 2012. Problem. HMM. Semantics. NLP Trinity. Parsing. Part of Speech Tagging. Morph Analysis.

keitha
Télécharger la présentation

CS626: NLP, Speech and the Web

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. CS626: NLP, Speech and the Web Pushpak BhattacharyyaCSE Dept., IIT Bombay Lecture 6, 7, 8, 9: Viterbi; forward and backward; Baum Welch; IL POS tags 6th, 8th, 9th, 13th August, 2012

  2. Problem HMM Semantics NLP Trinity Parsing PartofSpeech Tagging Morph Analysis Marathi French HMM Hindi English Language CRF MEMM Algorithm

  3. HMM Definition • Set of states: S where |S|=N • Start state S0 /*P(S0)=1*/ • Output Alphabet: O where |O|=M • Transition Probabilities: A= {aij} /*state i to state j*/ • Emission Probabilities : B= {bj(ok)} /*prob. of emitting or absorbing ok from state j*/ • Initial State Probabilities: Π={p1,p2,p3,…pN} • Each pi=P(o0=ε,Si|S0)

  4. Markov Processes • Properties • Limited Horizon: Given previous t states, a state i, is independent of preceding 0 to t-k+1 states. • P(Xt=i|Xt-1, Xt-2 ,…X0) = P(Xt=i|Xt-1, Xt-2… Xt-k) • Order k Markov process • Time invariance: (shown for k=1) • P(Xt=i|Xt-1=j) = P(X1=i|X0=j) …= P(Xn=i|Xn-1=j)

  5. Three basic problems (contd.) • Problem 1: Likelihood of a sequence • Forward Procedure • Backward Procedure • Problem 2: Best state sequence • Viterbi Algorithm • Problem 3: Re-estimation • Baum-Welch ( Forward-Backward Algorithm )

  6. Probabilistic Inference • O: Observation Sequence • S: State Sequence • Given O find S* where called Probabilistic Inference • Infer “Hidden” from “Observed” • How is this inference different from logical inference based on propositional or predicate calculus?

  7. Essentials of Hidden Markov Model 1. Markov + Naive Bayes 2. Uses both transition and observation probability 3. Effectively makes Hidden Markov Model a Finite State Machine (FSM) with probability

  8. Probability of Observation Sequence • Without any restriction, • Search space size= |S||O|

  9. Continuing with the Urn example Colored Ball choosing Urn 1 # of Red = 30 # of Green = 50 # of Blue = 20 Urn 3 # of Red =60 # of Green =10 # of Blue = 30 Urn 2 # of Red = 10 # of Green = 40 # of Blue = 50

  10. Example (contd.) Observation/output Probability Transition Probability Given : and Observation : RRGGBRGR What is the corresponding state sequence ?

  11. Diagrammatic representation (1/2) G, 0.5 R, 0.3 B, 0.2 0.3 0.3 0.1 U1 U3 R, 0.6 0.5 0.6 0.2 G, 0.1 B, 0.3 0.4 0.4 R, 0.1 U2 G, 0.4 B, 0.5 0.2

  12. Diagrammatic representation (2/2) R,0.03 G,0.05 B,0.02 R,0.18 G,0.03 B,0.09 R,0.15 G,0.25 B,0.10 R,0.18 G,0.03 B,0.09 U1 U3 R,0.02 G,0.08 B,0.10 R,0.06 G,0.24 B,0.30 R,0.24 G,0.04 B,0.12 R, 0.08 G, 0.20 B, 0.12 U2 R,0.02 G,0.08 B,0.10

  13. Probabilistic FSM (a1:0.3) (a1:0.1) (a2:0.4) (a1:0.3) S1 S2 (a2:0.2) (a1:0.2) (a2:0.2) (a2:0.3) The question here is: “what is the most likely state sequence given the output sequence seen”

  14. Developing the tree a2 a1 € 1.0 0.0 Start 0.3 0.2 0.1 0.3 • . • . 0.3 0.0 1*0.1=0.1 S1 S2 S2 S1 S2 S1 S2 S2 S1 S1 0.0 0.2 0.2 0.4 0.3 • . • . 0.1*0.2=0.02 0.1*0.4=0.04 0.3*0.3=0.09 0.3*0.2=0.06 Choose the winning sequence per state per iteration

  15. Tree structure contd… 0.09 0.06 a2 a1 0.3 0.2 0.1 0.3 • . • . 0.027 0.012 0.018 0.09*0.1=0.009 S2 S1 S1 S2 S2 S2 S1 S1 S1 S2 0.3 0.2 0.4 0.2 • . 0.0048 0.0081 0.0054 0.0024 The problem being addressed by this tree is a1-a2-a1-a2 is the output sequence and μ the model or the machine

  16. Tabular representation of the tree Latest symbol observed Ending state Note: Every cell records the winning probability ending in that state Final winner The bold faced values in each cell shows the sequence probability ending in that state. Going backward from final winner sequence which ends in state S2 (indicated By the 2ndtuple), we recover the sequence.

  17. Algorithm(following James Alan, Natural Language Understanding (2nd edition), Benjamin Cummins (pub.), 1995 Given: • The HMM, which means: • Start State: S1 • Alphabet: A = {a1, a2, … ap} • Set of States: S = {S1, S2, … SN} • Transition probability which is equal to • The output string a1a2…aM To find: The most likely sequence C1C2…CM which produces the given output sequence, i.e., C1C2…CM= argmaxC(P(C|a1a2…aM)

  18. Algorithm contd… Data Structure: • A N*M array called SEQSCORE to maintain the winner sequence always (N=#states, M=length of o/p sequence) • Another N*M array called BACKPTR to recover the path. Three distinct steps in the Viterbi implementation • Initialization • Iteration • Sequence Identification

  19. Initialization SEQSCORE(1,1)=1.0 BACKPTR(1,1)=0 For(i=2 to N) do SEQSCORE(i,1)=0.0 [expressing the fact that first state is S1] Iteration For(t=2 to T) do For(i=1 to N) do SEQSCORE(i,t) = Max(j=1,N) BACKPTR(I,t) = index j that gives the MAX above

  20. Seq. Identification C(M) = i that maximizes SEQSCORE(i,M) For i from (M-1) to 1 do C(i) = BACKPTR[C(i+1),(i+1)] Optimizations possible: • BACKPTR can be 1*M • SEQSCORE can be M*2 Homework:- Compare this with A*, Beam Search [Homework] Reason for this comparison: Both of them work for finding and recovering sequence

  21. Viterbi Algorithm for the Urn problem (first two symbols) S0 ε 0.2 0.5 0.3 U1 U2 U3 0.15 0.02 0.18 0.03 0.18 0.06 R 0.08 0.24 0.02 U1 U2 U3 U1 U2 U3 U1 U2 U3 0.015 0.04 0.075* 0.018 0.006 0.036* 0.048* 0.036 0.006 *: winner sequences

  22. Markov process of order>1 (say 2) • O0O1 O2 O3 O4 O5 O6 O7 O8 • Obs:ε R R G G B R G R • State: S0 S1 S2 S3 S4 S5 S6 S7 S8 S9 Same theory works P(S).P(O|S) = P(O0|S0).P(S1|S0). [P(O1|S1). P(S2|S1S0)]. [P(O2|S2). P(S3|S2S1)]. [P(O3|S3).P(S4|S3S2)]. [P(O4|S4).P(S5|S4S3)]. [P(O5|S5).P(S6|S5S4)]. [P(O6|S6).P(S7|S6S5)]. [P(O7|S7).P(S8|S7S6)]. [P(O8|S8).P(S9|S8S7)]. We introduce the states S0 and S9 as initial and final states respectively. • After S8 the next state is S9 with probability 1, i.e., P(S9|S8S7)=1 O0 is ε-transition

  23. Adjustments • Transition probability table will have tuples on rows and states on columns • Output probability table will remain the same • In the Viterbi tree, the Markov process will take effect from the 3rd input symbol (εRR) • There will be 27 leaves, out of which only 9 will remain • Sequences ending in same tupleswill be compared • Instead of U1, U2 and U3 • U1U1, U1U2, U1U3, U2U1, U2U2,U2U3, U3U1,U3U2,U3U3

  24. Forward and Backward Probability Calculation

  25. Forward probability F(k,i) • Define F(k,p)= Probability of being in state Si having seen o0o1o2…ok • F(k,i)=P(o0o1o2…ok , Sp) • With m as the length of the observed sequence • P(observed sequence)=P(o0o1o2..om) =Σp=0,N P(o0o1o2..om , Sp) =Σp=0,N F(m , p)

  26. Forward probability (contd.) F(k , q) = P(o0o1o2..ok , Sq) = P(o0o1o2..ok , Sq) = P(o0o1o2..ok-1 , ok ,Sq) = Σp=0,N P(o0o1o2..ok-1 , Sp , ok ,Sq) = Σp=0,N P(o0o1o2..ok-1 , Sp ). P(om ,Sq|o0o1o2..ok-1 , Sp) = Σp=0,N F(k-1,p). P(ok ,Sq|Sp) = Σp=0,N F(k-1,p). P(Sp  Sq) ok • O0O1O2O3 … Ok Ok+1 … Om-1Om • S0 S1S2 S3 … Sp Sq…SmSfinal

  27. Backward probability B(k,i) • Define B(k,i)= Probability of seeing okok+1ok+2…omgiven that the state was Si • B(k,i)=P(okok+1ok+2…om\ Si ) • With m as the length of the observed sequence • P(observed sequence)=P(o0o1o2..om) = P(o0o1o2..om| S0) =B(0,0)

  28. Backward probability (contd.) B(k , p) = P(okok+1ok+2…om \ Sp) = P(ok+1ok+2…om , ok |Sp) = Σq=0,N P(ok+1ok+2…om , ok , Sq|Sp) = Σq=0,N P(ok ,Sq|Sp) P(ok+1ok+2…om|ok ,Sq ,Sp ) = Σq=0,N P(ok+1ok+2…om|Sq). P(ok , Sq|Sp) = Σq=0,N B(k+1,q). P(Sp  Sq) ok • O0O1O2O3 … Ok Ok+1 … Om-1Om • S0 S1S2 S3 … Sp Sq…SmSfinal

  29. Assignment 2 • Prove that a language model based on POS tagged text is better than one developed from raw text LMpos > Lmraw • Choose a suitable NLP task to compare the models • Eg: next word prediction

  30. Example • Sentence: People laugh. • People and laugh can both be used as noun and verb • N V : P(“people laugh”, N V) = P(N).P(V|N).P(people|N).P(laugh|V) • N N : P(“people laugh”, N N) = P(N).P(N|N).P(people|N).P(laugh|N) • V N : P(“people laugh”, V N) = P(V).P(N|V).P(people|V).P(laugh|N) • V V : P(“people laugh”, V V) = P(V).P(V|V).P(people|V).P(laugh|V)

  31. Forward Probability • ^ People laugh $ • ^ N N $ V V P(People laugh) = F(People laugh, $) = F(People, N).P(N->laugh$)+F(people, V).P(V->laugh$)

  32. Label Sequence to Automaton N people laugh N people ε V laugh ^ $ laugh ε people N V laugh people V

  33. Forward Probability (contd.) • P(people laugh) = F(people laugh, $) = F(people, N).P(N->laugh$)+F(people, V).P(V->laugh$) • F(people, N)=F(^ people, N) =F(^, N).P(N->peopleN)+F(^, V).P(V->peopleN) • F(people, V)=F(^ people, V) =F(^, N).P(N->peopleV)+F(^, V).P(V->peopleV)

  34. HMM Training Baum Welch or Forward Backward Algorithm

  35. Key Intuition a Given: Training sequence Initialization: Probability values Compute: Pr (state seq | training seq) get expected count of transition compute rule probabilities Approach: Initialize the probabilities and recompute them… EM like approach a b a q r a b b b

  36. Baum-Welch algorithm: counts a, b String = abbaaabbbaaa Sequence of states with respect to input symbols a,b r q a,b a,b o/p seq State seq

  37. Calculating probabilities from table Table of counts T=#states A=#alphabet symbols Now if we have a non-deterministic transitions then multiple state seq possible for the given o/p seq (ref. to previous slide’s feature). Our aim is to find expected count through this.

  38. Interplay Between Two Equations wk No. of times the transitions sisj occurs in the string

  39. Illustration a:0.67 b:0.17 q r a:0.16 b:1.0 Actual (Desired) HMM a:0.04 b:0.48 q r a:0.48 b:1.0 Initial guess

  40. One run of Baum-Welch algorithm: string ababb State sequences * εis considered as starting and ending symbol of the input sequence string. Through multiple iterations the probability values will converge.

  41. Computational part (1/2) w0 w1 w2 wk wn-1wn S0  S1  S1  … Si  Sj …  Sn-1  Sn  Sn+1

  42. Computational part (2/2) w0 w1 w2 wkwn-1 wn S0  S1  S1  … Si  Sj …  Sn-1  Sn  Sn+1

  43. Discussions 1. Symmetry breaking: Example: Symmetry breaking leads to no change in initial values 2 Struck in Local maxima 3. Label bias problem Probabilities have to sum to 1. Values can rise at the cost of fall of values for others. a:0.5 b:0.25 a:0.25 s s s s s s b:1.0 a:0.5 a:0.5 b:0.5 a:1.0 b:0.5 a:0.25 b:0.5 b:0.5 Initialized Desired

  44. Indian Language POS tag standard

More Related